Towards Principled AI Alignment: An Evaluation and Augmentation of Inverse Constitutional AI
Undergraduate thesis advising for Esther An at Harvard SEAS. Topic focused on evaluating and extending inverse constitutional AI for principled alignment.
Projects
Selected work in AI alignment, machine learning, optimization, and multi-agent systems, including technical projects, open-source contributions, and student mentoring.
Mentoring
Research and course projects
Communication Complexity Bounds for Approximating Voting Rules
Final project for Harvard ECON 2070.
Expressivity of Shallow and Deep Neural Networks for Polynomial Approximation
Final project for MIT 9.520/6.860: Statistical Learning Theory and Applications.
Market Power and Strategic Drops: An Approach to Stable Matching Mechanisms
A study of market power and strategic participant withdrawals in stable matching mechanisms.
Edge of Stability Training Dynamics
Thoughts on EoS and self-stabilizing training dynamics in gradient descent, built as a final project for Harvard CS224.
Open-source and implementation
Contribution to Machine Learning Systems
I contributed to the open-source book Machine Learning Systems by Prof. Vijay Janapa Reddi.
Linear Mode Connectivity in Branching SGD Trajectories
An empirical study of linear mode connectivity between neural-network checkpoints produced by branching stochastic-gradient trajectories.
Accelerating LLM Decoding with Random Speculative Sampling
An implementation and experimental study of random speculative sampling for faster autoregressive language-model decoding.